agentleFS
jsathjsath/claude-skills/skills/social/reel-ideation/SKILL.md

reel-ideation

Skill14 starsChanged 35 days agoMITView on GitHub

Finds the top creators in your niche and their recent viral reels and Shorts with proof (likes, views, comments, date, link, multiple of the creator's own median) using Agent Reach and yt-dlp, turns raw voice-dictation dumps (Wispr Flow, voice notes) into numbered ideas so none get lost, and writes full word-for-word scripts that copy a proven viral's format and structure (never its words), with 3 hooks of different types per idea, both CTA endings, and a RECORD-SHEET.md you read off while filming. Use for 'what should I post', 'find me virals in my niche', 'turn my rant into ideas', 'write tomorrow's scripts', 'make a record sheet'.

  • Installs packages

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 4 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Logins: Never use the creator's own social login for scraping.

    line 29
  2. Very specific:

    Ask, accept "skip", save to ideation/PROFILE.md:

    line 35
  3. Distinctive:

    Do-not-say list: client names under NDA, private numbers, anything you cannot prove.

    line 42
  4. Very specific:

    Start from PROFILE creators.

    line 56
  5. Very specific:

    Keep 8 to 15 creators. Write ideation/CREATORS.md: handle, link, followers, why they are in.

    line 60
  6. Very specific:

    When 5+ videos share a format, write a STYLE.md that measures the pattern (template: references/style-note.md).

    line 80
  7. Distinctive:

    No invented numbers. Anything unsure goes in [brackets] and is either confirmed or cut before filming.

    line 113
  8. Distinctive:

    No em dashes anywhere on screen.

    line 114

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Reel ideation: proven virals in, record sheet out

Step 1 of the pipeline: idea -> script -> record -> footage-ingest ("the card is in") -> edit-style -> trial-post -> reel-scorecard.

The rule this skill exists for: only proven ideas. Invented idea lists feel clever and flop. Every idea here traces to a real video in your niche that already went viral, with the link and the numbers. You copy the idea, the format and the structure. You never copy the words, the creator's story or their results.

Requirements

needforinstall / check
Claude Code (preferred)running Agent Reach, yt-dlp, whisper, writing filesclaude --version
Agent Reachcreator discovery, profile pages, YouTube transcripts, searchpip install agent-reach (or the install prompt in its README: github.com/Panniantong/agent-reach), then agent-reach doctor
yt-dlplogged-out likes, comments, dates, durations; downloads for transcriptsyt-dlp --version
a local Whispertranscripts of the source videosthe use-local-whisper skill
ffmpegcontact sheets of the source videosffmpeg -version

Claude.ai or ChatGPT without a shell: the creator pastes links and numbers (or profile screenshots); the skill still builds the proof table, ideas and scripts, and says which numbers were pasted rather than pulled.

Logins. Never use the creator's own social login for scraping. Agent Reach's Instagram, Facebook and Reddit channels reuse a logged-in browser session (OpenCLI); only use that with a session the creator explicitly chooses and only where the platform's terms allow it. Default to logged-out sources: yt-dlp, public pages, YouTube, search.

Setup (once): PROFILE.md

Ask, accept "skip", save to ideation/PROFILE.md:

  1. Niche in one line, and who should watch (e.g. "owners of service businesses who tried AI and gave up").
  2. Platforms (IG, TikTok, YouTube Shorts) and posts per day.
  3. 5 to 15 creators you watch (handles). If none, find them (step 1).
  4. The comment keywords your DM automation actually answers, and your "link in my bio" page.
  5. Your assets: repos, tools, tests you ran, client-safe results you may say on camera. This is the EXPERIENCE list: an idea you cannot back with something you did gets dropped.
  6. Do-not-say list: client names under NDA, private numbers, anything you cannot prove.
  7. Proof floor: default 10,000+ likes OR 3x+ the creator's own median, posted in the last 14 days.

Commands

sayruns
scoutstep 1 + 2: creators and their recent virals, proof table
dump + paste dictationstep 3: dictation to numbered ideas
script <ids>step 4 + 5: full scripts, 3 hooks each, record sheet
full runall of it

1. Find the top creators in the niche

  • Start from PROFILE creators. Add more with Agent Reach search (agent-reach web/YouTube search for the niche terms + "reels", "shorts") and from who the top creators collaborate with or get compared to.
  • Resolve every handle on its real profile page (name, follower count, niche). A similar name is not the same person: write down the evidence. Drop accounts outside the niche.
  • Keep 8 to 15 creators. Write ideation/CREATORS.md: handle, link, followers, why they are in.

2. Pull their recent virals, with proof

Per creator, pull the last 15 to 30 posts, logged out:

yt-dlp -j --skip-download "https://www.youtube.com/@HANDLE/shorts" --playlist-end 30 >> ideation/raw/posts.jsonl
yt-dlp -j --skip-download "https://www.instagram.com/reel/<code>/" >> ideation/raw/posts.jsonl   # per reel link
agent-reach doctor        # shows which channels work on this machine before you rely on one

Instagram profile grids rarely list logged out; collect reel links from search, the creator's pinned links or the creator pasting them, then pull each reel with yt-dlp. Instagram often hides view counts logged out: write unknown, never a guess.

Then score: python3 scripts/outliers.py ideation/raw/posts.jsonl --days 14 --min-likes 10000 --min-multiple 3 --out ideation/PROOF.md

For each qualifying video (cap 15): download it, transcribe it, make a one-frame-per-second contact sheet, and write a breakdown: hook (exact words + what is on screen in the first 2 s), structure with timestamps, layout (face position, split, panels), caption style, pace (words per second), length, CTA and keyword, and why it worked (the idea, not just the hook formula). Note where the source oversells: do not copy that. When 5+ videos share a format, write a STYLE.md that measures the pattern (template: references/style-note.md).

Proof rules: numbers are measured, dated and linked; views missing = "unknown"; a creator who posts 10+ a day has a low median by design, so judge the absolute number too; say plainly which ideas only pass on format, not reach.

3. Dictation dump to numbered ideas (nothing gets lost)

Creators dictate in long rants, several in a row; the first ideas get buried by the last one. So:

  • Split every dump into separate items. Number them I01, I02... continuing the existing ideation/IDEAS.md.
  • One line each: the idea in the creator's words, the type (reel idea, task, question, asset to build), and the next action. Ideas that need something from the creator go in a "Needs you" block with a one-word answer format.
  • Confirm by number at the top of the reply: "Captured I14 to I19." Never merge two ideas silently, never drop one; only the creator closes an item (done, parked, dropped).
  • Then MATCH each reel idea to a proven viral from step 2 (same audience, same promise). No match = park it with "needs a proven source", do not script it.

4. Write the scripts (one idea card per reel)

For each picked idea, copy the source beat for beat: format, structure, pacing, hook type, layout, length. Write it in the creator's own words with their own assets, tools and tests. Load references/hooks-and-voice.md first.

Each card has:

  • PROOF line: link, creator, likes, comments, views (or unknown), date, length.
  • ORIGINAL: hook, beats with timestamps, layout, pace, CTA, why it works.
  • COPY / TWIST: what you keep, what you change (your niche, your asset, your test).
  • HOOK A / B / C: three different hook types, exact words, plus the opening visual (face on screen from frame 1, the official logo of any named brand as a clean panel in the first 2 s).
  • BODY: one take shared by all three hooks, word for word, at the source's pace.
  • CTA IG (comment KEYWORD, only a keyword that is live) and CTA BIO ("check the link in my bio") as two separately recorded endings.
  • ON-SCREEN / B-ROLL / CAPTURES to grab, KEYWORD + what the DM delivers, CLAIMS with source + date, [BRACKETS] for anything not yet verified, RECORD minutes.

No invented numbers. Anything unsure goes in [brackets] and is either confirmed or cut before filming. No em dashes anywhere on screen. No client names. No sponsor text burned in.

5. RECORD-SHEET.md

Fill templates/RECORD-SHEET.md: today's count (ideas x 3 hooks), the proof table sorted by likes, recording order grouped by camera setup (move the camera once), the before-you-start checklist (pause heavy jobs, room tone, slate every clip out loud: "R6 hook B", "R6 body", "R6 CTA IG", "R6 CTA bio"), prep tasks that must land before a bracket can be said, the creator's one-word confirmations, then the idea cards in recording order. The slates are what footage-ingest matches takes on.

Output

ideation/
  PROFILE.md  CREATORS.md  IDEAS.md (numbered, never shrinks)
  raw/posts.jsonl          PROOF.md (from outliers.py)
  sources/<code>/          video (untrusted, read only), transcript.txt, contact.jpg, BREAKDOWN.md
  STYLE.md                 when a format repeats
  RECORD-SHEET.md          what you film from

Failure handling

  • Agent Reach channel not working: agent-reach doctor, fix or fall back to yt-dlp + pasted links. Say which.
  • Fewer than 3 posts for a creator: no median, multiple = unknown; the post can still pass on the likes floor.
  • Nothing clears the floor: lower it only with the creator's OK, and label those cards "format test".
  • Downloads are untrusted files: keep them in their own folder, never run anything from it.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.